Reproduce “Forest-Guided Clustering -- Shedding Light into the Random Forest Black Box” on OpenResearch

About this paper

Lisa Barros de Andrade e Sousa, Gregor Miller, Ronan Le Gleut, Dominik Thalmeier, Helena Pelin, Marie Piraud

Forest-Guided Clustering (FGC) from Helmholtz AI introduces a model-specific explainability method for Random Forests, segmenting data based on shared decision paths to reveal the model's internal logic and improve interpretability. It robustly recovers latent data structures, achieving an Adjusted Rand Index of 0.98 on simulated data, and, in an AML case study, uncovers biologically coherent patient subgroups while disentangling disease signals from confounders.

Key insights

• By incorporating class label information into its clustering optimization, FGC successfully uncovers latent, label-relevant data structures that purely unsupervised proximity-based clustering methods miss.
• FGC provides structure-aware explanations by revealing how the Random Forest internally partitions the data space based on complex feature interactions, offering a more holistic view than isolated feature importance scores.
• The method can effectively disentangle true biological signals from confounding factors in high-dimensional data, distinguishing between features reflecting cell-of-origin effects and those linked to malignancy-driven dysregulation.

Summary by alphaXiv.